IEEE Transactions on Biomedical Engineering最新文献

A novel 3D-printed microwave probe operating in the 25-45 GHz frequency range is designed and fabricated for early skin tumor detection using signal processing. Due to the highly lossy nature of the skin, electromagnetic wave penetration is difficult. To overcome this limitation, a multi-section probe design was developed to enhance wave penetration into the skin layer. This design effectively mitigates the effects of high-loss tangents in tissues and compensates for the small size of tumors, aiding in early detection. The probe's performance is validated through simulations and experimental measurements, showing excellent agreement. For imaging evaluation, a phantom model composed of pork skin, measuring 30 mm × 30 mm with a skin thickness of 4 mm, is utilized. A total of 215 scanning points were analyzed, and time-domain reflection waves were extracted, demonstrating the probe's ability to detect variations in tissue properties accurately. These signals were then processed using an entropy-based method. The reconstructed images across various scenarios highlight the effectiveness of the proposed probe in achieving high-resolution microwave imaging, indicating its strong potential for non-invasive, early-stage tumor detection.

Restoring stiffness perception in prosthetic users through non-invasive methods remains a major challenge in haptic feedback research. This study evaluates a wearable vibrotactile stimulation system that conveys object stiffness information through two encoding strategies: Proposed Spatiotemporal Encoding and Circular Encoding, applied to two anatomical locations: upper-arm and forearm. Ten healthy participants completed structured trials of stiffness-discrimination involving vibrotactile cues corresponding to four stiffness categories. The results showed significantly higher classification accuracy (CA) and information transfer (IT) with the Proposed encoding strategy at both sites. The upper-arm-proposed configuration achieved peak performance (CA: 97.75%, IT: 1.84 bit/s), whereas the forearm-circular strategy yielded the lowest (CA: 73.62%, IT: 0.86 bit/s). NASA-TLX scores indicated a significantly lower mental workload for the proposed strategy, with the upper-arm feedback location providing superior perceptual clarity. A supplementary evaluation with a transradial amputee further demonstrated that the proposed encoding strategy remained interpretable, achieving classification accuracies above 85%. The classification accuracies over different conditions followed the same pattern as observed in healthy participants. These findings validate the importance of encoding geometry and stimulation site in designing effective haptic interfaces and support the feasibility of spatially distributed, non-invasive vibrotactile feedback for enhancing tactile perception in prosthetic applications.

Objective: Precise mealtime insulin bolus (MIB) dosing is essential in type 1 diabetes (T1D) to minimize glucose excursions from carbohydrate intake. Traditional MIB formulas, based on glucose concentration at mealtime, are suboptimal and do not exploit real-time data from continuous glucose monitoring (CGM). Existing methods that incorporate CGM data often rely on empirical rules or are developed in-silico, limiting their applicability to real-world conditions. This work investigates a framework combining machine learning (ML) algorithms, digital twins (DTs) and real-world data, to improve the assessment, tuning, and development of MIB dosing algorithms.

Methods: We utilized ReplayBG, a DT for T1D, to: i) evaluate a published linear ML model (Noaro et al.), originally developed in-silico, on real-world data from 30 free-living subjects; ii) recalibrate this model to fit the real-world dataset; and iii) train and test nonlinear gradient-boosting models (XGBoost, LightGBM) developed entirely on real data through DT simulations.

Results: Progressing to DT-enhanced models, we observed improvements in glucose control. The recalibrated linear and nonlinear models increased time-in-range (up to 80.6% with LightGBM vs. 75.6% for Noaro et al.) and reduced time-above-range. Risk metrics reflecting hypo/hyperglycemia also improved.

Conclusions: These findings demonstrate that a DT-based framework grounded in real-world data supports both the refinement and development of bolus calculators, achieving performance gains beyond the original in-silico model.

Significance: DTs allow the use real-world data to develop, validate and extend the domain of validity of new MIB formulas, paving the way to practical applications of ML tailored solutions for T1D.

Objective: Latent diffusion models (LDM) could alleviate data scarcity challenges affecting machine learning development for medical imaging. However, medical LDM strategies typically rely on short-prompt text encoders, nonmedical LDMs, or large data volumes. These strategies can limit performance and scientific accessibility. We propose a novel LDM conditioning approach to address these limitations.

Methods: We propose Class-Conditioned Efficient Large Language model Adapter (CCELLA), a novel dual-head conditioning approach that simultaneously conditions the LDM U-Net with free-text clinical reports and radiology classification. We also propose a data-efficient LDM pipeline centered around CCELLA and a proposed joint loss function. We first evaluate our method on 3D prostate MRI against state-of-the-art. We then augment a downstream classifier model training dataset with synthetic images from our method.

Results: Our method achieves a 3D FID score of 0.025 on a size-limited 3D prostate MRI dataset, significantly outperforming a recent foundation model with FID 0.070. When training a classifier for prostate cancer prediction, adding synthetic images generated by our method during training improves classifier accuracy from 69% to 74% and outperforms classifiers trained on images generated by prior state-of-the-art. Classifier training solely on our method's synthetic images achieved comparable performance to real image training.

Conclusion: We show that our method improved both synthetic image quality and downstream classifier performance using limited data and minimal human annotation.

Significance: The proposed CCELLA-centric pipeline enables radiology report and class-conditioned LDM training for high-quality medical image synthesis given limited data volume and human data annotation, improving LDM performance and scientific accessibility.

Objective: Achieving effective and robust free-living PD severity assessment with wearable intelligence technologies requires a deep understanding of clinically relevant features, representative activities, and machine learning algorithms.

Methods: We designed a unified analytic framework (PDWearML) to optimise wearable ML approaches with simple daily activities for fast assessment of PD severity. It comprises annotation criteria, feature importance analysis, representative activity combination, and PD severity assessment. We conducted a 12-month study, developing a supervised PD wearable dataset containing 100 PD patients and 35 age-matched healthy controls using Huawei smartwatches and Shimmer. PD severity, assessed by trained physicians using the Hoehn and Yahr (H&Y) scale.

Results: The results reveal that through optimising multi-level feature extraction and combining three representative daily activities (WALK, ARISING-FROM-CHAIR, and DRINK), our smartwatch-based machine learning approach can assess PD severity in supervised settings within 2 minutes with an accuracy of up to 84.7%.

Significance: This work holds significant clinical value, offering a potential auxiliary tool for faster, more tailored interventions in PD healthcare. Code is availableat code ocean platform and https://github.com/wang-xulong/PDWearML.

Variable-length time series classification (VTSC) problems are prevalent in healthcare applications, such as heart rate monitoring and electrophysiological recordings, where sequence length varies among patients and events. VTSC is challenging as finite-context models such as Transformers require padding, truncation, or interpolation, leading to distortion in the input data, higher computational costs, and overfitting, while infinite-context models including recurrent neural networks struggle with overcompression and unstable gradients over long sequences. In this paper, we develop a novel VTSC framework based on Stochastic Sparse Sampling (SSS) for seizure onset zone (SOZ) localization, a critical VTSC problem requiring identification of seizure-inducing brain regions from variable-length electrophysiological time series. The proposed framework sparsely samples time series windows to compute local predictions, which are then aggregated and calibrated to form a global prediction. SSS provides post-hoc insights into local signal characteristics related to the SOZ, by visualizing temporally averaged local predictions throughout the signal. We evaluate our method on the Epilepsy intracranial electroencephalography (iEEG) Multicenter Dataset, a heterogeneous collection of iEEG recordings obtained from four independent medical centers. The proposed solution outperforms state-of-the-art (SOTA) baselines across most medical centers, and superior performance on all out-of-distribution (OOD) unseen medical centers.

Identifying homogeneous subgroups with similar symptoms or neuropsychological patterns is essential for understanding the heterogeneity of psychotic disorders and advancing precision medicine, which enables tailored treatments based on patients' unique profiles. Existing data-driven methods, such as independent component analysis or independent vector analysis (ICA/IVA) applied to multi-subject functional magnetic resonance imaging (fMRI) data, have successfully revealed meaningful subgroups. However, these methods often rely on single-dimensional information, such as isolated functional networks, or assume uniform subgroup structures across all networks. Given the complexity of psychiatric disorders, exploring relationships across multiple functional networks can provide deeper insights into diagnostic heterogeneity. To address this, we propose a novel method that integrates cross-functional network information for subgroup identification by constructing multiplex networks from functional connectivity networks extracted from multi-subject resting-state fMRI data. Multiplex network-based community detection is then applied to identify both common communities spanning multiple networks and private communities specific to individual networks. Results from simulations and real-world fMRI data demonstrate the effectiveness of the proposed method. In a study of 464 psychotic patients, the identified subgroups exhibit significant differences in key functional areas, such as the default mode network (DMN) and anterior prefrontal cortex (antPFC), as well as corresponding clinical scores. These findings align with prior clinical studies, demonstrating the ability of the proposed approach to uncover clinically relevant subgroups and enhance understanding of psychotic disorder heterogeneity. By considering multi-dimensional information across functional networks, this approach provides a framework for understanding individual variability in psychotic disorders and paves the way for precision medicine.

Estimating the instantaneous phase of neural oscillations is crucial for technology that interfaces with the brain, such as brain-computer interfaces (BCIs) and neuromodulation systems. In these systems, phase information from the oscillating neural signal can be used to guide subsequent decisions to apply experimental perturbation. Traditional methods for phase estimation rely on the Hilbert transform computed using the Discrete Fourier Transform (DFT), which introduces a phase lag due to dependency on past and present signal values. This paper proposes a deep learning algorithm utilizing a dual-branch structure similar to the complex wavelet transform to generate a pseudo-complex valued signal for instantaneous phase estimation. The network has Discrete Cosine Transform (DCT) layers, which help to extract latent space representations for the real and imaginary signal components, respectively. An additional design goal was to make this Deep Learning (DL)-based algorithm suitable for deployment on portable edge devices with limited computing resources such as field-programmable gate arrays (FPGAs). This work demonstrates a proof-of-principle for real-time instantaneous phase estimation in neuromodulation applications. Our generalized model achieves an improvement of 40.3% in phase estimation accuracy over the endpoint-corrected Hilbert Transform (ecHT) method and an improvement of 9.2% over conventional deep learning model architectures.

Speaker identification in auditory attention decoding (SI-AAD) aims to identify the attended speaker from electroencephalography (EEG) signals. However, its application for hearing-impaired individuals is limited since existing methods rarely consider altered auditory perception from hearing-assistive devices under mixed-speech conditions, compounded by the lack of relevant datasets and difficulties in learning robust EEG-speech correspondences due to weak cross-modal alignment and insufficient feature extraction. Therefore, we construct five mixed-speech AAD datasets (MS-AAD), serving as the first EEG benchmark to simulate typical device-induced acoustic alterations without spatial cues. To enhance modality alignment, we propose a timbre-enhanced latent alignment (TELA) framework that jointly models latent embeddings and perceptual speaker cues via contrastive learning and auxiliary timbre classification. To further improve EEG-based feature extraction, we design FCTNet, a frequency-channel-temporal attention-based EEG encoder that captures rich neural patterns across multiple domains. Experiments on MS-AAD demonstrate that TELA and FCTNet jointly achieve 89.5% SI-AAD accuracy across diverse hearing conditions, highlighting the critical role of device-simulated acoustic dataset design and perceptually guided representation learning with advanced EEG encoding in mixed-speech SI-AAD for hearing-assistive applications.

Objective: Recent outbreaks and pandemics have emphasized the need for safe and reliable viral inactivation methods. The purpose of this work is to develop a simple and effective modeling approach to investigate viral inactivation via microwave absorption mediated by dipolar coupling.

Methods: Leveraging established techniques from the dynamic analysis of structures, a generalized Single-Degree-Of-Freedom (SDOF) model is developed, which is fully consistent with the dipolar resonance mode.

Results: The model can reproduce the main features of dipolar coupling with minimal computational time. Moreover, it allows mimicking the broadening of the resonance range associated with heterogeneous virus size, via Monte Carlo simulations, as well as water induced damping.

Conclusion: The results support the potential role of dipolar coupling for viral inactivation by microwave irradiation in the GHz range. The model can be used to assist in the interpretation of the experimental results, leading to an optimization of the inactivation protocols.

Significance: The proposed approach is versatile and can be extended to describe more complex cases, such as non-spherical geometries and/or non-homogeneous material properties.